Names, versions and owners change fast. Checked October 2026.
In plain words
AlphaGo is not a chatbot. It is a game-playing AI built by DeepMind, now part of Google DeepMind, for Go: an ancient board game where two players place black and white stones to surround territory. Go has about 10 to the power of 170 possible board positions, far too many for a computer to check one by one.
In March 2016, in Seoul, AlphaGo beat Lee Sedol, a player with 18 world titles, by four games to one. DeepMind says more than 200 million people watched the match.
AlphaGo learned in two stages. First it studied games played by strong human players. Then it played millions of games against itself and kept what worked, a method called reinforcement learning. A later version, AlphaGo Zero, skipped the human games entirely and became even stronger.
Everyday example
A strong carrom player planning a shot
A good carrom player does not try every possible shot. Experience tells them which three or four shots look promising. For each, they picture a few moves ahead and judge how good the board will look. Then they pick the best. AlphaGo does the same: one network suggests promising moves, another judges who is winning, and a search looks ahead along the best options.
Where the example breaks: A carrom player understands angles and physics. AlphaGo has no idea what a game is; it only learned which patterns of stones tend to lead to wins.
Latestas of October 2026
- Status
- Retired from competitive play after beating Ke Jie 3 to 0 in May 2017
- Successors
- AlphaGo Zero and AlphaZero (2017); DeepMind is now part of Google DeepMind
These details change often. The rest of this page explains what it is, which stays true.
How it actually works
AlphaGo combined two neural networks with a search method, as described in the journal Nature in January 2016.
The policy network: which moves to try
A network looks at the board and gives each possible move a probability of being good. It was first trained by supervised learning to predict the moves strong humans made in recorded games.
Learning from self-play
Copies of the network then played each other many times. Moves in won games were reinforced, moves in lost games discouraged. This reinforcement learning made it stronger than the humans it first copied.
The value network: who is winning
A second network learned to look at any position and estimate the chance of winning from it, trained on positions from those self-play games.
Tree search ties them together
During a game, AlphaGo used Monte Carlo tree search: it explored many possible future sequences, guided by the policy network toward promising moves and by the value network to judge where they lead, then played the most promising move.
Zero human knowledge
AlphaGo Zero, announced in October 2017, learned only by playing itself from random play. After three days of training it beat the version that defeated Lee Sedol by 100 games to 0.
- Made by
- DeepMind, London (now Google DeepMind)
- First win over a professional
- October 2015, 5 to 0 against European champion Fan Hui
- Lee Sedol match
- March 2016, Seoul; AlphaGo won 4 to 1
- Retired
- After beating Ke Jie 3 to 0 in May 2017
AlphaZero (2017) applied the same self-play idea to chess and shogi too. AlphaFold, which predicts protein shapes, is a different DeepMind system that does not play games.
Questions people ask
What was Move 37?
In game two against Lee Sedol, AlphaGo played a move that, by DeepMind's estimate, a human would choose only about 1 time in 10,000. It helped AlphaGo win the game. In game four, Lee's own rare Move 78 won him that game.
Did any human beat AlphaGo?
Yes, once. Lee Sedol won game four of their 2016 match. That was the only game a human won against AlphaGo in its official matches.
Is AlphaGo the same as AlphaFold?
No. Both come from DeepMind and both use deep learning, but AlphaGo plays Go, while AlphaFold predicts the 3D shapes of proteins. AlphaFold's creators shared the 2024 Nobel Prize in Chemistry.
Is AlphaGo a large language model?
No. AlphaGo works only with Go boards, not text. It shows a different branch of AI: learning by trial and reward. Similar reinforcement learning is now used to train reasoning models.