DeepMind is working with multiple game studios to design and prototype generalist agents and new gameplay experiences that aim to advance both interactive entertainment and AI research. The initiative targets systems that can operate in existing games without requiring changes to source code or special APIs.
Why games remain central to DeepMind
Since its founding in 2010, DeepMind has used the constrained yet rich environments of games to study intelligence. Early milestones include the Deep Q-Network (DQN), which learned to play 49 Atari 2600 games directly from pixels — including Pong, Breakout and Space Invaders — and whose 2015 Nature paper helped spark the modern era of deep reinforcement learning.
Subsequent achievements moved from board games to complex real-time strategy: AlphaGo defeated world champion Go player Lee Sae Dol in 2016; AlphaGo Zero learned solely through self-play without human data; AlphaZero generalized the approach to master chess, shogi and Go; MuZero learned to play without knowledge of the rules; and AlphaStar reached Grandmaster level in StarCraft II in 2019. Research in games also translated to other domains: foundations from game work contributed to AlphaFold and advances in protein structure prediction, recognized by the 2024 Nobel Prize in Chemistry.
From mastering games to understanding them
Past work showed AI can master games given clear objectives and enough training. DeepMind now asks a different question: can AI understand and interact with any game world as a person would?
The Scalable Instructable Multiworld Agent (SIMA) project addresses this. SIMA is not designed purely to maximize score; it is a generalist agent that ‘sees’ what a player sees on screen, understands natural language instructions, and acts using ordinary keyboard and mouse controls — no API or source code access required.
Built on Gemini frontier models, SIMA 2 functions as an interactive companion capable of real-time reasoning and conversation, achieving human-like play across complex 3D research environments and games such as No Man's Sky, Valheim and Hydroneer.
Potential benefits for developers and players
A truly general gaming agent could operate with existing games without code changes, enabling new gameplay forms: AI companions that genuinely understand the game world and NPCs that adapt and respond in ways scripted systems cannot. During development, such agents could provide robust QA as the game changes with each commit. Post-launch, they could adapt to new content or unexpected player behaviors in real time without re-scripting.
DeepMind develops SIMA with attention to safety and responsibility, partnering with established game studios and building a growing portfolio of games as research environments to challenge agents with increasingly complex tasks.
The Fenris Creations partnership and the EVE universe as a research platform
DeepMind’s major new research partnership is with Fenris Creations, the independent studio behind the EVE universe. Fenris Creations has spent more than two decades building a persistent multiplayer world. EVE Online, launched in 2003, is a single-shard universe where thousands of players share a continuously evolving game world with player-driven economies, supply-and-demand dynamics and trade networks across many star systems, shaped by alliances, conflict and diplomacy.
This living universe presents research opportunities for core frontier capabilities that DeepMind seeks:
- Continual learning: acquiring new skills without forgetting prior knowledge in a changing environment.
- Memory: accumulating and retrieving knowledge across timescales beyond current model context windows.
- Long-horizon planning: reasoning over weeks, months or years.
- Complex multi-agent dynamics: handling cooperation, competition, negotiation, economics and emergent social behavior at scale.
DeepMind’s program begins with an offline instance of EVE Online as a safe sandbox isolated from live players. Research can then progress into EVE Frontier — an environment with programmable “Smart Assemblies” and an extensible architecture suited to studying agents adapting to changing rules — and later consider EVE Online and EVE Vanguard only when capabilities are mature enough and the goal is to enrich human play.
EVE Vanguard, played from a first-person perspective, provides ground-level, fast-paced tactical decision-making within the broader persistent universe. Together, the EVE environments allow study of agents operating across multiple abstraction levels, from reflexive tactics to galaxy-scale strategy.
Early results and next steps
The collaboration has already produced practical player value: the Aura Guidance system uses Gemini to deliver player-generated knowledge and answers from Rookie Help to assist new pilots. DeepMind emphasizes a cautious rollout: starting in offline sandboxes and using staged environments before any integration into live services.
DeepMind reiterates its long-term aim to use AI as a catalyst rather than a replacement. The organisation seeks to unlock new gameplay experiences, improve accessibility and personalization in games, and apply lessons learned in games to real-world problems and scientific discovery, echoing past transitions from AlphaGo to AlphaFold.
DeepMind thanks its game partners for the collaboration and looks forward to sharing future discoveries.
Acknowledgements
DeepMind acknowledges many teams across Google DeepMind for their work on safe and responsible AI research in games. It gives special thanks to partner developers: Coffee Stain (Valheim, Satisfactory, Goat Simulator 3), Fenris Creations (EVE Online, EVE Vanguard, EVE Frontier), Foulball Hangover (Hydroneer), Hello Games (No Man's Sky), Keen Software House (Space Engineers), RubberbandGames (Wobbly Life), Strange Loop Games (Eco), Thunderful Games (ASKA, The Gunk, Steamworld Build), Digixart (Road 96) and Tuxedo Labs & Saber Interactive (Teardown).
