Google DeepMind Extends Game AI Research to EVE Online, Building on SIMA 2 Agent

Google DeepMind published a blog post outlining 15 years of AI research conducted through games, and announced an expanded research partnership with Fenris Creations, the independent studio behind EVE Online, to study AI capabilities in persistent multiplayer worlds.

From DQN to SIMA 2

The research line began in 2015, when DeepMind's Deep Q-Network (DQN) learned to play 49 Atari 2600 games directly from raw pixels, without game-specific engineering. The result was published in Nature and is widely credited with helping catalyze the modern deep reinforcement learning era.

Subsequent milestones, according to DeepMind, include:

• AlphaGo (2016): defeated world champion Go player Lee Sedol.
• AlphaGo Zero: surpassed prior AlphaGo versions using only self-play, no human data.
• AlphaZero: generalized the self-play approach to chess, shogi, and Go with a single algorithm.
• MuZero: learned to play without being given the rules of the game.
• AlphaStar (2019): reached Grandmaster level in StarCraft II, handling real-time play and imperfect information.

DeepMind says techniques developed for game-playing agents fed into AlphaFold, the protein structure prediction system that won the 2024 Nobel Prize in Chemistry.

The current focus has shifted from optimizing for high scores to building agents that understand game worlds the way people do. SIMA (Scalable Instructable Multiworld Agent) observes on-screen visuals, interprets natural language instructions, and acts via standard keyboard and mouse input, without API or source-code access to the game. SIMA 2, built on DeepMind's Gemini models, is described by the company as capable of real-time reasoning and conversation, achieving human-like play across 3D environments including No Man's Sky, Valheim, and Hydroneer.

New partnership with EVE Online's developer

DeepMind is expanding its research collaboration with Fenris Creations, the studio behind EVE Online, a massively multiplayer space simulation that launched in 2003 and has run continuously for more than 20 years with a player-driven economy spanning thousands of star systems.

According to DeepMind, EVE Online's persistent, player-shaped world presents four research challenges central to its broader agenda:

• Continual learning — acquiring new skills without forgetting prior ones, in a constantly changing environment
• Memory — retaining and retrieving knowledge across timescales beyond current model context windows
• Long-horizon planning — reasoning over weeks, months, or years
• Multi-agent dynamics — cooperation, competition, negotiation, and emergent social behavior at scale

The partnership extends to Fenris Creations' broader EVE Universe, including EVE Vanguard, a first-person tactical variant of the same persistent world, and EVE Frontier, offering additional environments for studying agents across different levels of abstraction — from moment-to-moment tactics to galaxy-spanning strategy.

Fenris Creations CEO Hilmar Pétursson said the collaboration pushes into "uncharted territory where AI must learn, adapt and remember on timescales that no other game environment demands," framing it as a way to study human-AI coexistence in a virtual setting before comparable questions arise in the real world.

DeepMind also cited existing collaborations with studios including Hello Games, Coffee Stain Studios, and Foulball Hangover, describing a "show, don't tell" approach that pairs its research team with developers to prototype new gameplay rather than retrofit existing AI techniques onto games.

No specific product timeline, technical benchmarks, or performance metrics for the EVE Online work were disclosed.

What this means

This is a research and partnership update, not a model release — no new model, weights, or version was announced. The substantive claim worth tracking is the shift in research target: from games with clear win conditions (Go, chess, StarCraft) to persistent, open-ended worlds with no score to optimize, where success is harder to measure and verify. EVE Online's two-decade player-driven economy makes it a genuinely distinctive testbed for long-horizon memory and multi-agent behavior, but DeepMind has not published benchmarks, timelines, or technical specifics for what the SIMA line of agents will actually be asked to do inside it. Until concrete results appear, this should be read as a signal of research direction rather than a demonstrated capability.