DeepMind EVE Online: What the AI Will Learn

DeepMind and Fenris Creations partner on EVE Online

DeepMind calls its research partnership with Fenris Creations a new chapter in its history of AI research in games. The partnership is built around EVE Online.

Fenris Creations is the independent studio behind EVE

DeepMind's post describes Fenris Creations as independent, and it quotes the studio's CEO, Hilmar Pétursson, on what the research is meant to explore. The post says the research extends across Fenris Creations' expanding universe, which offers distinct environments for AI development.

EVE Online has run as a single shared universe since 2003

EVE Online is a space game that went live in 2003. Thousands of players share one world, and that world has kept changing for more than 20 years. In the post's words, it is a single-shard persistent universe: every player is in the same copy of the game, and it carries on from one session to the next.

EVE Vanguard adds first-person tactics

EVE Vanguard is the second game in the partnership. It is played from a first-person perspective, so an AI agent there faces fast, ground-level tactical decisions inside the same persistent world that EVE Online provides at large scale.

Four capabilities are the stated research targets

The post names continual learning, memory, long-horizon planning and multi-agent dynamics. The table gives each one in the post's own terms.

The four capabilities named in the announcement

Capability	What it covers
Continual learning	Acquiring new skills without forgetting earlier ones, in a world that keeps changing
Memory	Storing and retrieving knowledge over timescales far beyond today's model context windows
Long-horizon planning	Reasoning over weeks, months or even years
Multi-agent dynamics	Cooperation, competition, negotiation, economics and emergent social behavior at scale

Memory beyond context windows, and continual learning

A context window is the amount of text a model can look at in one go. The post says the memory target reaches far past that limit, to knowledge gathered and recalled over long stretches of time. Continual learning is the companion problem: an agent in a world that never stops changing has to pick up new skills and keep the old ones.

Long-horizon planning spans weeks, months or years

The post defines long-horizon planning as reasoning over weeks, months or even years. EVE Online has kept running for more than 20 years.

Multi-agent dynamics cover cooperation and negotiation

The fourth target is about many agents acting in one world. The post lists cooperation, competition, negotiation, economics and emergent social behavior, all at scale. Pétursson adds a second purpose: learning how humans and AI can coexist in a virtual environment before the same questions come up in real life.

SIMA 2 plays via the screen, language and game controls

SIMA is DeepMind's agent for playing games the way a person does: it looks at the screen, follows language instructions, and presses keys. SIMA 2 runs on Gemini and is described as capable of real-time reasoning and conversation. DeepMind's model for controlling robots is covered in Gemini Robotics 2: Whole-Body and Multi-Robot.

SIMA needs no API or source-code access

For game developers, the post says, a general gaming agent would work with existing games, with no changes to the game code. SIMA works from what a player sees and from ordinary keyboard and mouse input, so a game does not have to expose an API or share its source code.

The partnership extends DeepMind game research from 2010

DeepMind has used games to study intelligence since it was founded in 2010. The research began with a network that learned Atari 2600 games from raw pixels.

That network, the Deep Q-Network, learned 49 games with no game-specific engineering. Its 2015 Nature paper helped start modern deep reinforcement learning, a method in which a program learns by trial and error from rewards.

After that came AlphaGo, which beat a world champion Go player in 2016, and AlphaZero, which used one algorithm for chess, shogi and Go. AlphaStar reached Grandmaster level in StarCraft II in 2019. The post also credits the same foundations for AlphaFold, which tackled the 50-year challenge of predicting protein structures and was recognized with the 2024 Nobel Prize in Chemistry.

The announcement gives no date, benchmark or product

The post describes a research partnership. It does not set a release date, publish a benchmark or promise a product or feature for EVE Online players. What it does fix is the list of two games and four capabilities above.