Google DeepMind Brings Gemini AI Research to EVE Online
Google DeepMind and Fenris Creations will test Gemini powered agents across EVE games, pursuing adaptive play, stronger QA and long term AI research program.
Summary
Google DeepMind’s August 21, 2026 report by Alexandre Moufarek and Adrian Bolton reviews research since its 2010 founding with Demis Hassabis. DQN learned 49 Atari 2600 games, including Pong, Breakout and Space Invaders, from pixels; its 2015 Nature paper spurred deep reinforcement learning. AlphaGo beat Lee Sae Dol in 2016 and produced Move 37; self taught AlphaGo Zero preceded AlphaZero for chess, shogi and Go, and rule blind MuZero. AlphaStar reached StarCraft II Grandmaster in 2019. The approach led to AlphaFold’s solution of the 50 year protein structure challenge and 2024 Nobel Prize in Chemistry.
Gemini powered SIMA 2 interprets screens and instructions, operates keyboard and mouse without APIs or source code, and plays, reasons and converses in No Man’s Sky, Valheim and Hydroneer. DeepMind targets adaptive companions and NPCs, code independent QA and post launch adaptation. Its Fenris Creations partnership covers EVE Online, EVE Vanguard and EVE Frontier. Online, launched in 2003, offers a 20 plus year single shard universe with thousands of players and star systems, economics, diplomacy and conflict for studying continual learning, memory, long horizon planning and multi agent dynamics. Vanguard adds first person tactics; Frontier adds programmable Smart Assemblies and mutable rules. CEO Hilmar Pétursson links the work to human AI coexistence.
Aura Guidance uses Gemini and Rookie Help material for new pilots. Testing starts offline in EVE Online, advances through EVE Frontier, then may reach live Online and Vanguard when mature. Partners and games also include Coffee Stain, Satisfactory and Goat Simulator 3; Foulball Hangover; Hello Games; Keen Software House and Space Engineers; RubberbandGames and Wobbly Life; Strange Loop Games and Eco; Thunderful Games, ASKA, The Gunk and Steamworld Build; Digixart and Road 96; Tuxedo Labs, Saber Interactive and Teardown. A November 2025 post details SIMA 2. DeepMind seeks accessible, personalized play and scientific transfer, not human replacement.
Positives
- DQN mastered 49 Atari 2600 games from pixels, helping its 2015 Nature paper accelerate deep reinforcement learning.
- SIMA 2 uses ordinary keyboard and mouse controls without APIs or source code across No Man’s Sky, Valheim and Hydroneer.
- Aura Guidance already applies Gemini and real Rookie Help material to assist new EVE Online pilots.
- EVE Online, EVE Vanguard and EVE Frontier provide environments spanning immediate tactics, persistent economics and galaxy scale strategy.
- Offline EVE Online testing precedes any possible deployment with live players, creating a phased research path.
Risks & concerns
- EVE research requires memory beyond current model context windows and planning across weeks, months or years.
- Live use in EVE Online and EVE Vanguard will be considered only after capabilities mature, leaving deployment timing uncertain.
- EVE Frontier’s mutable mechanics demand agents that can adapt when the rules themselves change.
- A truly general gaming agent remains a research objective rather than a capability demonstrated across every existing game.
- DeepMind provides no numerical EVE agent benchmarks or schedule for progressing beyond the offline sandbox.