SIMA 2 is the latest iteration of DeepMind's AI agent, building on its predecessor, SIMA. While the original SIMA could follow over 600 basic commands in various video games, SIMA 2 transitions from merely following instructions to becoming an interactive gaming companion. It integrates the advanced reasoning capabilities of the Gemini models, allowing it to understand high-level goals, communicate with users, and improve its performance over time. The training process for SIMA 2 involved a mix of human demonstrations and Gemini-generated labels. This has enabled the agent to grasp more complex instructions and respond effectively, even in unfamiliar gaming scenarios like the Viking survival game ASKA. A standout feature is its ability to generalize learning across different tasks, such as applying the concept of “mining” in one game to “harvesting” in another. This flexibility brings SIMA 2's performance closer to human players, showcasing its adaptability in completely new environments generated by another project, Genie 3. SIMA 2 also demonstrates a capacity for self-improvement. Initially trained with human input, it can later learn through self-directed play, refining its skills without needing further human data. Despite these advancements, SIMA 2 faces challenges with complex, multi-step tasks and has limitations in memory management and precise action execution. The research serves as a validation for action-oriented AI, suggesting that broad competency can unify various specialized systems into a cohesive generalist agent. The team aims to develop SIMA 2 responsibly, engaging with academics and game developers for feedback as they explore its potential applications.