Google DeepMind has introduced SIMA 2, the next evolution of its generalist AI agent that operates in virtual 3D worlds. Moving beyond basic instruction-following like its predecessor, SIMA 2 now functions as an interactive gaming companion capable of reasoning about goals, conversing with users, and improving its own performance over time through experience.

By integrating Gemini models as its core architecture, SIMA 2 understands complex, nuanced instructions across diverse gaming environments including ASKA and MineDojo. It demonstrates remarkable generalization by transferring concepts between games and can describe its intentions and reasoning process. The agent even operates effectively in newly generated virtual worlds created by Genie 3.

Perhaps most impressive is SIMA 2's self-improvement capability, allowing it to learn through trial-and-error without constant human intervention. This represents a significant step toward Artificial General Intelligence with applications in robotics. Despite limitations with complex multi-step tasks, SIMA 2 demonstrates that broad, multi-world training combined with advanced reasoning can create coherent generalist agents.