This is the paper describing AlphaGo, published in Nature volume 529, pages 484–489, on 27 January 2016 (doi:10.1038/nature16961). The authors are David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, Nal Kalchbrenner, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel and Demis Hassabis, of what was then Google DeepMind; Silver and Huang are noted as contributing equally, with Silver and Hassabis as corresponding authors.
Method
The paper frames Go as the hardest of the classic games for artificial intelligence, on account of its search space and the difficulty of evaluating board positions and moves. Its approach combines three components:
- Value networks to evaluate board positions.
- Policy networks to select moves.
- A search algorithm combining Monte Carlo simulation with both networks.
The networks are trained by what the paper describes as a novel combination of supervised learning from human expert games and reinforcement learning from games of self-play. The authors report that without any lookahead search, the networks alone play at the level of state-of-the-art Monte Carlo tree search programs that simulate thousands of random self-play games.
Reported results
With the combined search algorithm, AlphaGo achieved a 99.8% winning rate against other Go programs and defeated the human European Go champion 5 games to 0. The paper states this was the first time a computer program had defeated a human professional in the full-sized game of Go, "a feat previously thought to be at least a decade away."
Position in the record
The paper is the second of the two results conventionally used to date the deep-learning era's early phase, after AlexNet. Its methodological contribution — self-play reinforcement learning producing superhuman performance in a domain where supervised imitation of human play plateaus below the human ceiling — became a recurring reference point in later arguments about capability trajectories, and its "a decade away" framing became a standard citation in the forecasting-error literature. The DeepMind team it came from is now Google DeepMind.
Provenance
Retrieved July 31, 2026 from nature.com using the vault's bin/fetch-source.py; Firecrawl was rate-limited during this cycle. The abstract, full author list, publication date, volume, page range, and DOI were captured from the article page. The full text is behind a subscription wall, so this page summarizes the abstract, citation metadata, and the reference list, not the methods and results sections.
Relationships
- related: ImageNet Classification with Deep Convolutional Neural Networks (Krizhevsky, Sutskever and Hinton, NeurIPS 2012) — the other conventionally cited opening result of the deep-learning era.
- related: Google DeepMind, Demis Hassabis — the originating laboratory and a corresponding author.
- depends-on: RLHF (Reinforcement Learning from Human Feedback) — self-play reinforcement learning is the antecedent of the RL post-training methods used on language models.
- related: AGI Timelines — the paper's own note that the result was thought "at least a decade away" is a standard datapoint on expert forecasting error.