Where does AlphaGo go now?

Originating in China more than 3,000 years ago, the game Go has a simple objective: Surround the other player’s stones with your own. However, the complexity lies in the near infinite combination of moves that can be played on a grid-like board with 361 intersections (and total stones). Because of those limitless possibilities, AI researchers and engineers have long used the game to study AI .

Go’s intricacies, decision-making, and logic and strategic thinking confounded AI researchers for decades. “Go is considered one of the grand problems of artificial intelligence. It’s a problem that seems like it would be easy to crack, but it’s not,” said Peter Drake, then an assistant professor of computer science at Lewis and Clark College, in 2004. Fast forward almost 20 years, and science has caught up, to the point where a loss rings alarm bells.

In March 2016, after beating 18-time world champ Go player Lee Sedol in three consecutive games, the AI system known as AlphaGo had a hiccup.

In the fourth game of the five-game series, set at the Four Seasons Hotel in Seoul, Lee used his 78th move of the match to wedge a white token between his AI opponent’s two black ones, which appeared to put the machine on tilt. Lee won Game 4.

Lead researcher of AlphaGo David Silver cancelled activities for the next day. There would be no sightseeing tour until they figured out if their machine was hallucinating.

The team set their sights instead on bugs in AlphaGo and came up empty. “The bug was that Lee Sedol came up with an ingenuous move,” team member Ioannis Antonoglou wrote to us. AlphaGo beat Lee in the final match, making the final tally: robots: 4, human professional: 1.

Since that matchup, AlphaGo has racked up Ws in chess and its Japanese form, shogi. Now, almost a decade later, the large language models that Sliver and team whiteboarded, then deployed, have led to possibilities beyond game night, including various applications for both consumer and corporate users.

According to a recent Gallup workplace study, 44% of surveyed US employees in white collar industries say their organization has begun integrating AI to generate ideas (41%), consolidate information (39%), and automate basic tasks (39%).

Silver, now VP of reinforcement learning at Alphabet-owned AI company DeepMind, and Antonoglou, reflected on building AlphaGo and where it goes from Go. And as AI flips stones, what comes next as we move from an ancient board game to the beginning of a new technological era?

“We’ve moved from a simple game with simple rules to the real world,” Silver said.

Let’s AlphaGoooo. In the months leading up to the AlphaGo match, Silver, then a research scientist at DeepMind, spent lots of time in front of a whiteboard. There he explored ideas with a core team that, at its max, reached about 12 people—a mix of engineers, researchers, Go players, and hardware efficiency experts like Antonoglou, who had to figure out how to get the models to run effectively on the GPUs and then brand-new TPUs.

Some of those whiteboard sessions: trying a different neural network, maybe one that reflects structures and symmetries common to the game; then running the code and keeping track of scores on an internal leaderboard.

“Whenever you try to just do something at that scale, you need to just really be focused and really just be dedicated, and you just do whatever it takes,” Antonoglou told us.

According to Google, the AlphaGo project began in 2013, with Silver and fellow DeepMind researcher Aja Huang. In 2014, Chris Maddison, then being advised by Google Brain’s Ilya Sutskever, co-inventor of neural network AlexNet, joined the team. Over the next two years, the team expanded and in 2014 was acquired by Google.

“You’re coming up with these algorithms and then seeing them do incredible things for the first time. Just watching these lines of code turn into something which can rival the human brain in this particular area of expertise, it’s a magical process to go through, and we had a lot of fun doing it,” Silver said.

Bring in the reinforcements! AlphaGo relies on two neural networks: The “policy network” makes a move, and the “value network” predicts the game’s winner.

Testing AlphaGo meant introducing the systems to numerous amateur games and playing it thousands of times against itself— learning from mistakes in a process known as reinforcement learning.

“If it wins the game, it reinforces the kinds of things which it did in that game. If it loses the game, it does the opposite, and makes those things less likely,” he said.

The AlphaGo team spent about two years on the project, Silver told us, before the matchup against Lee. While there was no pressure to work 100-hour weeks, he said, it was hard to stop an engaged team from putting in the effort and time.

The matchup. Once the Go champ agreed to the match in Seoul, “things became real,” Silver said, and the team focused on getting the technology ready for the five-game series, scheduled for March 9, 2016. These preparations were documented in the 2017 film, AlphaGo.

The machine won the first three rounds handily, which clearly had a discouraging effect on Lee. At the end of the second match, he said to reporters: “Yesterday I was surprised, but today I’m quite speechless.”

Game 2 also featured a move from AlphaGo that felt like a mistake to announcers and many of the Go players in the room. At Move 37 of the second match, DeepMind’s Huang, manually playing AlphaGo’s decided moves, placed a black stone in the middle of empty space on the board’s fifth line—an unconventional choice that AlphaGo determined only 1 in 10,000 human players would have played in that instance.

“It went beyond its human guide and it came up with something new, and creative, and different,” Silver said in the documentary. The move surprised Lee, too, who looked puzzled after coming back from a smoke break and seeing the stone. That 37th placement, however, offered a game-deciding connecting point that ultimately led AlphaGo to victory.

“AlphaGo had come in as a massive underdog. And then there was this moment where everything just turned on its head, and people just started to see the game in a new way and realize that something special was happening,” Silver told us.

AlphaGo, AlphaGoing, AlphaGone. For Lee, the post-series feeling appeared complicated—somewhere between encouragement and defeat.

In the film, he said he felt grateful for the experience, which affirmed why he plays the game. “I’ve grown through this experience. I will make something out of it with the lessons I have learned,” he said. His unexpected, machine-tricking Game 4 move that appeared to put AlphaGo on tilt is often referred to as a “God’s Touch.”

Now retired from the game, Lee said to the New York Times at an education fair in Seoul that he “faced the issues of AI early,” and that “it may not be a happy ending,” referring to the technology’s potential impact.

An April poll from Pew Research Center found that 56% of AI researchers believe the technology “will have a very or somewhat positive impact” on the US, while only 17% of non-expert US adults felt that way.

Where AlphaWent. An upgraded version of AlphaGo known as “AlphaGo Zero,” according to research published in December 2017, ​​achieved (within 24 hours) a “superhuman level” of play in chess and shogi. The “zero” in AlphaGo Zero refers to the level of domain knowledge beforehand, beyond basic game rules. DeepMind models AlphaProof and AlphaGeometry 2 have solved International Mathematical Olympiad problems.

DeepMind’s algorithms, Silver said, have moved from a closed game to the real world and its complex scientific challenges, like developing nuclear fusion to solve the energy crisis or discovering cures for diseases.

Antonoglou, now CTO and co-founder of autonomous coding agent maker Reflection, notes the continuing popularity today of games like Go and chess, and how powerful AI can unlock human productivity and ingenuity.

“Having access to something that’s better than us in particular tasks actually helps us be better in other things that we care about,” he said, “to either learn from them or to actually use them to build more powerful machines or more powerful tools.”

In 2025, AI users have evolved from small teams of 12 trying to figure out how to beat the world’s best chess and Go players to, well, everybody. Consumers are adopting ChatGPT for everyday use, and enterprise pros are charged with implementing AI tools responsibly—meaning companies and IT teams still have a lot of work to do.

Gallup’s 2025 poll found that only 22% of surveyed enterprise AI users said their organization has communicated a clear plan or strategy for implementing the technology.

“We are bringing things into the world which have powerful capabilities, but I think the right way to do this is to think of a collaboration where we are trying to build tools that can actually help humans to be more productive, to do the things that they want to do more effectively,” Silver said. “If we can really push this forward right and make this happen the right way, I think we’ll end up with really exciting progress in technology.”

For now, we wait to see which way we—and AI—go.

This is one of the stories of our Quarter Century Project, which highlights the various ways industry has changed over the last 25 years. Check back each month for new pieces in this series and explore our timeline featuring the ongoing series.

Visit the timeline