August 21, 2026 Research
Alexandre Moufarek and Adrian Bolton
From Atari to Go and StarCraft, games have driven some of the most significant breakthroughs in artificial intelligence. Today, we are working with game developers to prototype new kinds of gaming experiences and advance the shared evolution of games and AI.
Since DeepMind was founded in 2010, the constrained yet richly varied worlds created by games have played a vital role in advancing our understanding of intelligence. From mastering Atari games to helping solve the problem of protein structure prediction, games have driven some of our biggest breakthroughs in AI—and they remain at the heart of our work today.
Games are part of GDM’s DNA. One of Google DeepMind’s founders, Demis Hassabis, was himself a game developer, and many members of the GDM team have similar backgrounds. Together, we bring decades of experience in game development and a deep respect for the craft of making games.
We have always believed that using games for AI research requires deep collaboration with game developers. This includes the significant research partnership with Fenris Creations announced earlier this year, our work in the EVE universe, and our collaborations with leading studios such as Hello Games, Coffee Stain Studios, and Foulball Hangover.
Games: An engine for AI research
Our exploration began with a small team that trained deep neural networks to play Atari 2600 games directly from raw pixels. The Deep Q-Network (DQN) learned to play 49 different games—from Pong and Breakout to Space Invaders—without any game-specific engineering or optimization. The 2015 Nature paper on DQN helped usher in the modern era of deep reinforcement learning.
We then sought to master increasingly complex games, with each milestone bringing more capable and generalizable systems. In 2016, AlphaGo defeated world Go champion Lee Sedol—an achievement many experts had believed was at least a decade away. AlphaGo Zero learned entirely through self-play, without using any human data, surpassing all previous versions. AlphaZero generalized this approach to chess, shogi, and Go, mastering all three games with a single algorithm. And MuZero learned to play games without even knowing their rules. In 2019, AlphaStar reached Grandmaster level in StarCraft II, handling real-time complexity and imperfect information.
Across each of these games, AI enriched the player experience. AlphaGo’s famous move 37 was so unexpected that professional commentators initially thought it was a mistake. The move overturned centuries of conventional wisdom in the Go world and inspired experts to explore new strategies. AlphaZero similarly inspired entirely new lines of play in chess. More importantly, the spirit of exploration that led to success in games had a profound impact on other AI systems: AlphaFold applied these foundations to solving the longstanding 50-year challenge of protein structure prediction. This breakthrough was ultimately recognized with the 2024 Nobel Prize in Chemistry.
From mastering games to understanding them
Our early work demonstrated that, given a clear objective and sufficient training, AI could master any game. But the real world has no scoreboards or rulebooks. This led us to a fundamentally different question: Could AI understand and interact with any game world as people do?
That is the challenge SIMA is designed to address. SIMA stands for Scalable Instructable Multiworld Agent. Rather than optimizing for a high score, SIMA is a general-purpose agent: it can “see” what the player sees on screen, understand natural-language instructions, and act using a standard keyboard and mouse—without using APIs or accessing source code.
Powered by Gemini frontier AI models, SIMA 2 can act as an interactive companion, reasoning and conversing in real time. It can demonstrate human-like gameplay in complex 3D research environments and video games, including No Man’s Sky, Valheim, and Hydroneer.
For game developers, a truly general game agent would unlock AI capabilities for existing games without requiring changes to the game code. This could enable entirely new forms of gameplay, such as AI companions that genuinely understand the game world and non-player characters (NPCs) that adapt and respond in ways that scripting systems have never been able to achieve.
General-purpose game agents could also transform how games are made. During development, games change with every code commit, and these agents could enable genuinely robust quality-assurance (QA) testing. Once a game is released, they could adapt in real time as new content is introduced or players behave unpredictably, generalizing to new situations without requiring scripts to be rewritten.
To develop SIMA agents safely and responsibly, we are collaborating with leading game studios and gradually building an expanding portfolio of games for AI research. This allows us to challenge agents with increasingly complex tasks—capabilities that may eventually transfer to solving problems in the real world.
Collaborating with game developers to explore new frontiers in AI and games research
Game studios bring expert craftsmanship, extraordinary game worlds, and a deep understanding of players. We bring frontier AI—from Gemini to research into generative interactive environments and embodied agents—as well as research expertise, the team’s unique backgrounds in game development, and years of experience building AI for interactive environments. Together, we are focused on discovering breakthrough experiences and creating entirely new forms of gameplay that would only be possible with AI.
The focus is not the technology itself, but the fun of the experience. That is why we take a “show, don’t tell” approach with our partners. Our teams work side by side with game developers to explore new ideas, build playable prototypes, and find the experiences that are genuinely fun.
Our latest research collaboration with Fenris Creations marks a new chapter in our history of AI research in games. Fenris Creations is the independent studio behind the EVE universe.
For more than two decades, Fenris Creations has been building one of the game industry’s most extraordinary persistent worlds. Launched in 2003, EVE Online is a massively multiplayer space simulation in which thousands of players share a single universe that has continued to evolve for more than 20 years. Its player-driven economy features real supply and demand, along with trade networks spanning thousands of star systems. Shaped by alliances, conflict, and diplomacy, this world is driven by interactions between people.
For AI research, this presents an exceptional opportunity. It is a continuously operating, ever-evolving world that demands precisely the capabilities we believe are essential to frontier AI:
- Continual learning: Acquiring new skills in a constantly changing world without forgetting what has already been learned.
- Memory: Accumulating and retrieving knowledge across timescales far beyond the reach of today’s model context windows.
- Long-horizon planning: Reasoning over the course of weeks, months, or even years.
- Complex multi-agent dynamics: Navigating cooperation, competition, negotiation, economic activity, and emergent social behavior in large-scale environments.
These challenges lie at the heart of our broader research program, which aims to build systems that can continually learn from experience and improve their rate of learning over time. We believe these frontier capabilities could unlock