Exploring new frontiers of AI and games research — Google DeepMind

Google DeepMind traces its AI research roots to games like Atari, where systems such as the Deep Q-Network learned to play multiple titles directly from pixels without game-specific modifications. This approach later expanded to complex strategy games, with milestones including AlphaGo’s 2016 victory over Lee Sae Dol and AlphaZero’s mastery of chess, shogi, and Go using a single algorithm. These breakthroughs not only advanced gaming but also contributed to broader scientific progress, such as AlphaFold’s protein structure prediction, recognized with the 2024 Nobel Prize in Chemistry.

The company’s latest initiative, SIMA (Scalable Instructable Multiworld Agent), shifts focus from high-score optimization to creating AI agents that interact with games as humans would. Powered by the Gemini model, SIMA 2 processes on-screen visuals, follows natural language instructions, and uses standard controls to play games like No Man’s Sky and Valheim without requiring game code access. This generalist approach aims to enable AI companions and adaptive NPCs that respond dynamically to player behavior and in-game events.

For developers, SIMA offers tools to enhance game creation and testing, including robust QA during development and real-time adaptation to new content or player actions post-launch. Google DeepMind emphasizes collaboration with studios like Fenris Creations, Hello Games, and Coffee Stain Studios to prototype AI-driven gameplay experiences. These partnerships blend frontier AI research with game development expertise to explore novel interactions that traditional scripting cannot achieve.

A recent collaboration with Fenris Creations, the studio behind EVE Online, highlights the potential of persistent, evolving game worlds for AI research. EVE Online’s player-driven economy and continuous evolution over 20 years present challenges in long-term learning and adaptation, mirroring real-world complexities. The partnership extends to other EVE titles, including EVE Vanguard and EVE Frontier, which offer varied environments for studying multi-level decision-making and adaptive mechanics, further bridging gaming innovation and real-world AI applications.