AlphaGo: AI at the Movies
Note- this contains plenty of spoilers. If you want to watch AlphaGo before reading this, you can watch it for FREE on YouTube.
Man vs. Machine
“It’s not that they (Google’s DeepMind) want to understand Go. They want to understand what understanding is, and maybe that is truly what it means to be human.” — Frank Lantz in AlphaGo.
In March 2016, a momentous event in the history of AI occurred. And it was all captured on film. AlphaGo, an AI system developed by Google’s DeepMind, took on Lee Sedol, the world’s greatest Go player, in a series of matches that would reverberate through technology and geopolitics for years to come. The documentary AlphaGo beautifully tells this story as an archetypal Man vs. Machine conflict.
Lee Sedol, a 9 dan player (the highest recognition in Go), would play a five-game series against AlphaGo. Most commentators, including Lee, assumed he would win all the games. In 2016, Go was considered too complex for computers to master. Calculating all the potential moves like a chess computer would is mathematically impossible for Go. (This was being discussed as far back as 1961 by mathematician I. J. Good, “So I think it will be even more difficult to programme a computer to play a reasonable game of Go than of chess.”) Humans play Go through intuition. And computers don’t have intuition. Or so we believed.
The Match
The DeepMind team went to South Korea, where Go is not just a game but a deeply ingrained part of the culture. With a history spanning over 1500 years, it is revered as an art form and a symbol of intellectual prowess. The significance of the match was not lost on the millions of Koreans who tuned in to watch the live broadcasts.
Game 1 shocked the world as AlphaGo triumphed, leaving Sedol visibly shaken. Sedol entered the second match, determined to prove his loss was a fluke. However, AlphaGo won yet again and proved that it was no fluke.
Move 37
In game two, AlphaGo made a move that many commentators and experts initially thought was a mistake. It was Move 37 of the match. Even the DeepMind team was flummoxed. The movie took us to that moment when a panicked team ran the probabilities, thinking AlphaGo might have made an error. There was only a 1/10,000 chance an expert human player would have made that move.
Needing a moment, Lee Sedol stepped out for a smoke break. When he returned and saw AlphaGo’s move, he was taken aback. It took him over twelve minutes to make his next move. While the rest of the Go world thought the machine was failing, Lee saw something else (after all, he is the greatest Go player in the world). He later commented, “Sure, AlphaGo is creative. This move was really creative and beautiful. This move made me think about Go in a new light.”
To put that into perspective, Lee is the greatest Go player. He’s been playing since he was eight and has been a professional since he was twelve. But this algorithm, this non-breathing entity, did something so atypical and unique that it changed how he thought about the game he devoted his life to.
Not all was lost for humanity as Lee found a way to beat AlphaGo in the fourth game, giving humanity one opportunity to celebrate. But the AI won again in game five. This series made it clear that no human would again be able to beat an AI in Go, and the game itself would be forever changed.
(For further reading, enjoy this Quora post by Corrin Lakeland explaining the uniqueness of Move 37 for a Go-playing novice and this CNN piece on How AI turned the ancient sport of Go upside down.)
China vs the US
Chinese culture has a similar affinity to Go, as does Korea. In many ways, witnessing this event “woke China up” to the power of artificial intelligence. The following year, China’s State Council published its Generation Artificial Intelligence Development Plan, setting goals and outlining the strategic importance to the country in AI. Last May, former Google CEO Eric Schmidt told a US House committee the following:
I am here to discuss the foremost strategic issue of our time: the technological aspect of the competition between the United States and the People’s Republic of China… 280 million Chinese watched the AI program AlphaGo play the Chinese strategy game Go and defeat the best Go player in the world. I know the shock it sent to the Chinese people — because I was there in person for that game. Since then, China has dedicated enormous resources in an effort to outpace the United States in those technologies I consider fundamental to this strategic competition.
The History and Future of DeepMind
DeepMind was founded by Demis Hassibis, Shane Legg, and Mustafa Suleyman in 2010. (Suleyman was not featured in the movie and is now the CEO of Microsoft AI.) Facebook was looking to acquire the company, but Google outmaneuvered the Facebook team and bought it in 2014, allowing DeepMind to run autonomously within Google. DeepMind’s team focused on building models to play games using reinforcement learning. Effectively, they would tell the model very little about a game and let the model play with a singular goal of “winning” the game. In this example, you can see how DeepMind’s model eventually learns to “beat” the game by playing the Atari game Breakout.
This reinforcement learning formed the building blocks for some of the most exciting opportunities in AI today. One of them is AlphaFold 3, which was announced last month. The model can help scientists predict the shapes and interactions of important proteins and other molecules, better understand diseases and develop new medicines.
It is no secret that AI engineers are a highly coveted commodity. Those who have worked in the same AI company for ten-plus years are a rare breed. I looked up all the DeepMind employees prominently featured in the film and found that five out of seven at least appear to be still working at Google/DeepMind today. So many are still there, which is quite a testament to the culture that they’ve built.
5 Fast Facts About AlphaGo
- Lee Sedol retired from Go soon after playing AlphaGo. “With the debut of AI in Go games, I’ve realized that I’m not at the top even if I become the number one through frantic efforts,” he said.
- The techniques used to train AlphaGo, particularly reinforcement learning and self-play, have been applied to other complex games like chess and shogi (Japanese chess). DeepMind built AlphaZero in 2017 and mastered those games with just 24 hours of training.
- DeepMind has owned the rights to the film and has chosen to publish it on YouTube, a Google subsidiary, since March 2020.
- In April 2023, DeepMind merged with the other major AI lab inside of Google, Google Brain, to form a single AI research organization for the company.
- AlphaGo has an honorary 9 dan professional rating from the South Korean Go Association on the eve of the fifth match against Lee Sedol. This is the only “honorary” rating the association has ever granted.
Next Time
Come back in two weeks, June 24, for our third film for AI At the Movies: 2001. It is perhaps the most famous film about an AI gone wrong. For folks in the US, you can stream it on Max, which is also available in many international markets.
