{
  "id": 2375076,
  "title": "From Atari to EVE Online: Building on 15 Years of AI Research in Games",
  "url": "https://urgent.news/2026/08/21/from-atari-to-eve-online-building-on-15-years-of-ai-research-in-games",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-21T11:59:48.000Z",
  "source": {
    "name": "Google DeepMind",
    "slug": "google-deepmind",
    "url": "https://deepmind.google/blog/from-atari-to-eve-online-building-on-15-years-of-ai-research-in-games/"
  },
  "original_language": "en",
  "account": "For more than two decades, Fenris Creations has meticulously crafted EVE Online, a sprawling space simulation that reflects the consequences of real-world supply and demand. At its core, this persistent universe is shaped by the intricate dance of human alliances, conflicts, and diplomacy. As a game, EVE Online has become an unparalleled testbed for pioneering artificial intelligence research. DeepMind, the UK-based artificial intelligence company where Demis Hassabis and other founders played as game developers, has been at the forefront of this partnership since 2019.\n\nInitially, DeepMind's Deep Q-Network (DQN) demonstrated the potential of AI in games, mastering 49 Atari 2600 games from raw pixel input. This breakthrough catalyzed the modern era of deep reinforcement learning. One of DQN's most significant accomplishments was defeating world champion Go player Lee Sae Dol in 2016 with AlphaGo, a feat many experts thought would take a decade to achieve. Building upon this success, AlphaGo Zero and AlphaZero took AI to new heights by learning entirely from self-play, without any human data. AlphaZero then went a step further by mastering chess, shogi, and Go with a single algorithm.\n\nRecently, AlphaStar reached Grandmaster level in StarCraft II, showcasing the ability of AI to navigate complex, real-time environments and imperfect information. These advances in AI have enriched the gaming experience, with AlphaGo's iconic Move 37 leaving professional commentators baffled and inspiring new strategies in Go. The spirit of exploration that propelled these AI achievements has also impacted other AI systems, such as AlphaFold, which applied the same foundations to protein structure prediction and earned the 2024 Nobel Prize in Chemistry.\n\nWhile games provide a controlled environment for AI to master, the real world is far more unpredictable and complex. This led to the development of SIMA, a Scalable Instructable Multiworld Agent designed to understand and interact with game worlds as a human would. Powered by Gemini, a frontier AI model, SIMA 2 can see, comprehend natural language instructions, and act through standard keyboard and mouse controls — all without requiring access to game APIs or source code. SIMA 2 demonstrates human-like play across complex 3D environments like No Man's Sky, Valheim, and Hydroneer.\n\nThe potential of a general gaming agent goes beyond entertainment. Such an agent could revolutionize game development by enabling robust Quality Assurance testing and adapting to new content in real-time as it is introduced. By partnering with game studios like Fenris Creations, DeepMind brings frontier AI research and game development expertise together, focusing on creating breakthrough experiences that push the boundaries of gaming and AI.",
  "summary": "Google DeepMind partners with game studios to prototype breakthrough AI gameplay.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}