Google DeepMind teams up with EVE Online maker to test AI that plays like a person
The lab behind AlphaGo wants game characters that actually understand what's happening on screen. Its new partner? The studio running a 22-year-old space universe.
Key points
- Google DeepMind has announced a research partnership with Fenris Creations, the independent studio behind EVE Online, which launched in 2003.
- The lab's generalist gaming agent SIMA 2, built on its Gemini AI models, follows spoken instructions and plays titles like No Man's Sky and Valheim using only a keyboard and mouse.
- DeepMind's 2015 Deep Q-Network learned 49 Atari games from raw pixels; AlphaGo beat world champion Lee Sedol in 2016.
- Those game-research foundations later fed into AlphaFold, recognised with the 2024 Nobel Prize in Chemistry for predicting protein structures.
- EVE Online's continuously running universe will be used to test AI skills including long-term memory and planning across weeks or months.
Google DeepMind, the AI lab inside Google that built AlphaGo, has announced a research partnership with Fenris Creations, the studio behind the long-running space game EVE Online.
The pitch is simple. DeepMind wants AI characters that behave less like scripted robots and more like curious players. Fenris has a 22-year-old online universe full of real humans doing real strategy. Put them together and you've got a giant sandbox for testing smarter AI.
What is DeepMind actually building?
The headline project is SIMA 2, short for Scalable Instructable Multiworld Agent. It's a generalist gaming agent: one piece of software that can play many different games without being retrained for each one.
SIMA 2 doesn't plug into a game's code. It looks at the screen the way you do, listens to instructions in plain English, and clicks the mouse and taps the keyboard. No special access required. It runs on Gemini, Google's family of large AI models, and DeepMind says it can already hold a conversation while playing titles including No Man's Sky, Valheim, and Hydroneer.
Why partner with a 22-year-old space game?
Because EVE Online is genuinely weird, in a useful way. Launched in 2003, it puts thousands of players into a single shared galaxy that never resets. Alliances form, wars happen, and a player-run economy trades goods across thousands of star systems.
That's exactly the kind of messy, long-running environment where today's AI tends to fall over. DeepMind lists four skills it wants to test there:
| Capability | What it means in plain English |
|---|---|
| Continual learning | Picking up new skills without forgetting old ones |
| Memory | Remembering things for weeks, not just one chat |
| Long-horizon planning | Thinking months ahead, not just the next move |
| Multi-agent dynamics | Cooperating, competing and negotiating with others |
The partnership also covers EVE Vanguard, a first-person shooter set in the same universe, and EVE Frontier, an experimental version where players can reprogram the rules of the world itself.
How did we get here?
Games have been DeepMind's testing ground since the beginning. In 2015 the lab's Deep Q-Network learned to play 49 Atari 2600 games, from Pong to Space Invaders, just by staring at pixels. The following year AlphaGo beat world champion Lee Sedol at Go, a feat many experts thought was still a decade away. AlphaZero later mastered chess and shogi with a single algorithm, and AlphaStar reached Grandmaster level at StarCraft II in 2019.
Those same techniques then jumped out of games entirely. AlphaFold used them to predict the shapes of proteins, work recognised with the 2024 Nobel Prize in Chemistry.
None of that happened in a leadership vacuum. We reported in August that Demis Hassabis has stepped back from day-to-day leadership at DeepMind, handing the reins to Koray Kavukcuoglu. This EVE partnership is one of the first major research moves announced under that transition.
What does this mean for players?
Nothing ships tomorrow. This is research, not a product launch, and DeepMind hasn't promised any of it will appear inside EVE Online.
But the direction is clear. A generalist agent that drops into an existing game without touching its code could serve studios two ways: AI testers that hammer a build after every update, and non-player characters (the shopkeepers and quest-givers you meet throughout a game) that actually respond to what you do rather than repeating the same lines.
One privacy point worth raising before any of this reaches live games: SIMA agents work by watching the screen, so a future rollout would need clear rules about what an AI companion sees and stores about your play. That's the question to ask before you invite one into your save file.
