Google DeepMind announces SIMA 2, an AI that plays games like a human

From high scores to world understanding

Google DeepMind has summarized fifteen years of AI research in video games. Evolving from their initial breakthroughs with Atari 2600 games in 2015, the team is now focusing on SIMA (Scalable Instructable Multiworld Agent). The new SIMA 2 generation no longer optimizes for high scores using internal game rules. Instead, the agent learns to understand the environment exactly as a human player does: reading the screen output and controlling the keyboard and mouse based on text instructions.

The end of custom integrations

For game studios, this represents a fundamental shift in how they can integrate AI. SIMA 2 does not require modifying the game's source code or building custom API endpoints. Powered by Gemini models, the system acts as a universal player capable of completing tasks in complex 3D environments like No Man's Sky or Valheim. DeepMind is currently prototyping new gameplay experiences in partnership with Fenris Creations and the EVE Universe.

The limits of a universal player

This approach comes with a clear trade-off. While screen reading and keyboard simulation remove the need for custom integration, they introduce significant computational latency. The agent is currently better suited for strategic tasks with clear instructions rather than reacting in real-time to fast-paced, unpredictable events in competitive multiplayer settings.

Adoption beyond the lab will decide

Google is demonstrating that the gaming industry is not just a testing ground for other fields, but a production environment in its own right. The true test for SIMA 2 won't be how well it plays old games, but whether developers will start building entirely new game mechanics specifically designed around such agents.

Lilith's verdict

Google doesn't need a better No Man's Sky player. It needs to sell developers an agent they can integrate without touching their own source code.