From Atari to EVE Online: Building on 15 Years of AI Research in Games

Since its founding in 2010, Google DeepMind has used games as a testbed for AI research, citing co-founder Demis Hassabis's own game development background. The lab argues that games provide constrained but rich environments and have driven breakthroughs from Atari to protein structure prediction. DeepMind emphasizes that partnerships with game developers are essential, pointing to its new research collaboration with Fenris Creations and the EVE Universe, as well as work with Hello Games, Coffee Stain Studios, and Foulball Hangover.

The research journey began with the Deep Q-Network (DQN), a deep neural network trained on raw pixels that learned to play 49 Atari 2600 games without game-specific engineering, and its 2015 Nature paper catalyzed deep reinforcement learning. Subsequent milestones include AlphaGo defeating Lee Sedol in 2016 (a feat many experts thought was a decade away), AlphaGo Zero learning entirely from self-play, AlphaZero mastering chess, shogi, and Go with a single algorithm, MuZero learning without knowing the rules, and AlphaStar reaching Grandmaster level in StarCraft II in 2019.

These systems also influenced human play: AlphaGo's 'Move 37' appeared to be a mistake to experts but overturned centuries of Go wisdom, and AlphaZero inspired new chess strategies. The same exploration principles led to AlphaFold, which solved the 50-year protein structure prediction challenge and was recognized with the 2024 Nobel Prize in Chemistry.

DeepMind now says earlier work showed AI could master any game given a clear objective, but the real world doesn't come with such objectives. The lab is partnering with game developers to prototype new gameplay experiences that push the frontiers of both gaming and AI, moving from mastering games to understanding and applying AI to real-world complexity.