From Atari to EVE Online: Building on 15 Years of AI Research in Games
DeepMind says it’s using games to test new AI, and it’s now working with EVE’s studio Fenris Creations. The twist: it wants agents that can learn, plan, and adapt inside live game worlds, not just chase scores.
Based on reporting by Google — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind is going back to the place where a lot of its AI story began: games. On August 21, 2026, the company said it is partnering with game developers to prototype new gameplay experiences, with Fenris Creations and the EVE Universe now part of that effort.
The pitch is bigger than making bots that play well. DeepMind says games have been central to its work since 2010, from early Atari research to later systems that helped push forward protein structure prediction. The company’s argument is simple: games are constrained enough to study, but rich enough to expose real intelligence problems.
That history runs through a string of headline systems. DQN learned to play 49 Atari 2600 games from raw pixels. AlphaGo beat Lee Sae Dol in 2016, AlphaGo Zero learned from self-play alone, AlphaZero moved across chess, shogi and Go, MuZero learned without knowing the rules, and AlphaStar reached Grandmaster level in StarCraft II in 2019. DeepMind also points to AlphaFold, which drew on that same exploratory approach.
The newer focus is less about scorekeeping and more about understanding worlds. SIMA, its Scalable Instructable Multiworld Agent, is meant to see the screen like a player, follow natural-language instructions, and operate with keyboard and mouse controls. SIMA 2, powered by Gemini, is described as an interactive companion that can reason and talk in real time, with human-like play across environments including No Man’s Sky, Valheim and Hydroneer.
Fenris Creations gives DeepMind a much messier test bed. EVE Online has run since 2003 as a single shared universe with a player-driven economy and thousands of star systems. The company says that kind of environment is useful because it forces agents to deal with continual learning, memory, long-horizon planning and complex multi-agent behavior. EVE Vanguard and EVE Frontier add first-person tactics and a programmable, open-ended world on top of that.
DeepMind says the collaboration has already produced one player-facing result: Aura Guidance, which uses Gemini to surface knowledge from Rookie Help questions and answers for new pilots. For everything else, the plan starts offline in a sandbox, then moves toward studying how humans and agents coexist, and only later — if the systems are mature enough — into live games.
My take — AI-written commentary, not fact-checked reporting
This is the sensible version of AI-in-games: start with messy worlds, keep the live players out of the blast radius, and don’t pretend a chatbot with a health bar is a revolution. The real story is that game studios are now being asked to provide the proving ground for frontier AI, which is exactly the kind of arrangement that sounds neat until the bot starts “learning” in production.
Read more about this at: Google